Goodbye Claude, Hello Pi

I’ve been a long time Claude Code user, since it appeared in the Claude Pro membership strategy over a year back. I keep in mind the very first time I utilized it, feeling giddy as it spat out untidy Python code for whatever random weekend task I was dealing with, all from a single timely. I’m particular that in a couple of years we’ll all recall at its release as one of the significant chauffeurs of this brand-new period of software application advancement we’re in.

Regardless of this, due to a build-up of lots of little aggravations, I’ve lastly chosen to retire it as my primary coding harness and move my advancement setup to utilize Pi rather.

Somebody at Anthropic need to’ve captured wind … simply a day after making my choice, they revealed Claude Code would lastly begin supporting AGENTS.md files in an effort to get me to remain. For me it was too little, too late.

Grumbling

In lieu of an appropriate obituary, I wish to take this time to state a couple of words about the important things I liked least about Claude Code (and therefore what I’m anticipating preventing with Pi).

The AGENTS.md thing type of characterizes the entire factor I chose to stop utilizing Claude Code. Do not get me incorrect, having a lots of CLAUDE.md files spread about numerous GitHub pastime repos is fantastic guerrilla marketing. It draws for users to have to deal with copying or symlinking in between the 2 formats.

Like the Luddite I am, I still choose to evaluate the majority of the code it produces before I devote it. I utilize(d) Claude Code inside VSCode, and it lets you link the harness to the IDE. This is good since you can highlight text in a file and have it instantly consisted of in the context when asking concerns or asking for modifications throughout an evaluation.

Other than you can just have one session “linked” to the IDE at a time. If you open a 2nd tab to begin another job while the very first one works, it can’t get the text choice occasions from the IDE at all. This has actually been reported at least as far back as December 2025… and closed as a replicate of an absolutely unassociated bug.1

Naturally, you can constantly utilize the main Claude Code VSCode GUI extension, which does enable you to have several tabs open and load the text choice into the context of each. This extension has plenty of issues of its own. Half of the integrated slash commands do not operate at all in the GUI. It does not have any assistance for a customized status line. Subagent output messages are merely layered into the primary chat thread in sequential order, which is a mess to check out when more than one is running. I might continue.

In spite of all of these existing concerns, it looks like Anthropic has actually operated at a feverish rate to load more functions into Claude Code without much look after quality or stability. Dynamic workflowsrepresentative groupsultrareviews — all fantastic methods to assist business consumers burn ever more tokens every day, however not actually anything I desire or require.

Together with all of these function releases, they constantly appear to be adjusting things in manner ins which screw up my individual setup. At the time of composing, there have actually been 244 releases to Claude Code this year; over 3/4 of these altered the system timely or tool descriptions, which can significantly alter how the LLM carries out (specifically when these brand-new guidelines oppose user triggers or repo conventions).

And what are all those modifications for? In spite of slicing 80%of the system timelyClaude Code still has method more overhead than Pi for relatively no advantage. The default Bash tool’s description is nearly 3,000 tokens, the majority of which is random things about utilizing git and opening PRs on GitHub.2 The designs currently understand how to do this without in-depth guidelines, by the method.3

Looking Forward

This post has actually been a great deal of disliking. I’m sorry! That’s not my normal design, and I do believe Anthropic has actually made a truly cool piece of software application here. I’m likewise not under any deception that I would’ve done a much better task if I were the one in charge of running Claude Code from the start. LLM coding harnesses are a brand-new area and being among the very first to innovate in a field is an actually difficult job.

Later on tools like Pi have the advantage of hindsight to see what works and what does not, what’s a core function and what can be cut.

Eventually, CC seems like an item without a strong vision or instructions. It’s as if a lot of thrilled engineers and PMs are quickly adding whatever functions fit their workflow … however without factor to consider for how this may make things even worse for other users. It looks like there isn’t somebody high-up whose task is strictly to state “no” to these demands, which is extremely essential when it concerns item engineering. For now, I’m leaving.

I’m delighted to lean more into Pi. The very first couple of days needed some finagling, however ever since it’s been smooth cruising. I’m taking pleasure in the stability, minimalist technique, and the severe level of customizability. I’ve currently composed a number of extensions to support my usage cases, which I’m intending to share right after I’ve cleaned them up a bit. If you’re likewise feeling these disappointments, I ‘d extremely motivate providing an open source harness a shot!